Cluster Ensembles Based on Vector Space Embeddings of Graphs

Abstract

Cluster ensembles provide us with a versatile alternative to individual clustering algorithms. In structural pattern recognition, however, cluster ensembles have been rarely studied. In the present paper a general methodology for creating structural cluster ensembles is proposed. Our representation formalism is based on graphs and includes strings and trees as special cases. The basic idea of our approach is to view the dissimilarities of an input graph g to a number of prototype graphs as a vectorial description of g. Randomized prototype selection offers a convenient possibility to generate m different vector sets out of the same graph set. Applying any available clustering algorithm to these vector sets results in a cluster ensemble with m clusterings which can then be combined with an appropriate consensus function. In several experiments conducted on different graph sets, the cluster ensemble shows superior performance over two single clustering procedures.